Decision Theory and Discourse Particles: A Case Study from a Large
Japanese Sentiment Corpus
Christopher Davis
University of Massachusetts Amherst
Abstract. The distribution and use of the Japanese particle yo is examined using a large annotated sentiment corpus. The data is shown to support a decision-theoretic account ofyo’s meaning (Davis, 2009). A decision-theoretic approach to the analysis of sentiment corpora is proposed, by which empirical predictions of decision-theoretic formal analyses can be tested using large sets of naturalistic data.
Keywords: discourse particles, sentiment corpora, corpus pragmatics, decision theory, Japanese
1 Introduction
There has been a recent surge of interest in the formal semantics and pragmatics literature on the topic ofdiscourse particles(Zimmermann, to appear). Discourse particles straddle the border of semantics and pragmatics, and provide a perfect empirical domain for developing and challenging formal models of linguistic meaning. Discourse particles are, as the name implies, connected to the context of an entire discourse, and force the analyst to go above the sentence level and develop a theory of discourse contexts within which sentences and their associated particles are situated and interpreted.
One problem for the development of formal theories of discourse particles is the fact that they typically make no truth-conditional contribution to the sentences in which they occur, and the con-tribution that they do make is typically very difficult to pin down. Most formal studies of particles rely on intuitionistic data from small sets of typically constructed examples. The ineffability and extreme context-sensitivity of discourse particles make it difficult to study them using corpora and other naturalistic data, in which the analysist is unable to control the discourse context and cannot probe the often subtle speaker intuitions that guide the use of these particles.
Recently, a number of researchers have exploited large sentiment corpora to explore empirical regularities in the use ofexpressivesand other emotionally-charged language (Potts and Schwarz, 2008; Constant et al., 2008; Davis and Potts, to appear). The structure of these corpora has al-lowed researchers to explore the use of these often ineffable items using large sets of naturalistic texts, on the basis of which empirical estimates of the expressive effects of this kind of language can be made. In this paper, I expand on this line of research by showing how sentiment corpora can be used as an empirical tool in the exploration ofdecision-theoreticanalyses of the semantics and pragmatics of lexical items and constructions. I focus on a particular formal analysis of the Japanese sentence final discourse particle yo (Davis, 2009). This analysis builds on recent de-velopments in decision/game-theoretic semantics and pragmatics (Parikh, 2001; van Rooy, 2003; Benzet al., 2005a). By testing the formal analysis with quantitative data from naturalistic texts, I demonstrate the utility of corpus methods for lexical pragmatics.
pro-to the common ground ofc(Stalnaker, 1978), where the common ground of a contextc,CG(c), is modeled as a set of propositions.
The denotation of yo takes a context change potential (CCP) and context (c) as arguments, returning a (pragmatic) presupposition. This presupposition relies on a set of contextually salient actions, A(c0), representing the options from which our agent(s) must choose.2 Formally, the alternative actions are understood asproperties, so that for a given worldw, actiona, and agent
x, we havea(x)(w) = 1iffx choosesainw, anda(x)(w) = 0iff xdoes not chooseainw. In the presupposition associated withyo, we require that there is some action in this set3 which is optimalin the output contextc0, but isnot optimal in the pre-update contextc. Optimality is defined as follows:
(4) Definition of Optimality:
OPT(p,c) = 1 iff∀wi, wj ∈ ∩CG(c) [(p(wi) ∧ wi <c wj)→p(wj)]
A propositionpis optimal relative to a contextcjust in casepis true of all the most highly-ranked worlds consistent with the CG inc, where ranking is determined by the contextual ordering defined in (5).
(5) Partial Ordering of Worlds [modified from (Portner, 2007)] For all worldswi, wj ∈ ∩CG(c),wi<c wj iff
∃p∈g(c) [p(wj) &¬p(wi) &∀q∈g(c) [q(wi)→q(wj) ] ], whereg(c)is theordering sourceinc. (Kratzer, 1981)
The contextualordering sourceis a set of propositions, upon which the partial order<cis defined. A worldwi is ordered above another worldwj if there is a proposition in the ordering source that holds ofwiand not ofwj, while all other propositions in the ordering source that hold ofwj also hold ofwi.
While the above semantics may seem a bit complicated, its consequences for pragmatics are easily summarized. According to (2), the use of yopresupposes that there is some actionathat is salient in the post-update context and is also optimal for the addressee in that context, but not optimal in the pre-update context. Putting the pieces together, what this means is that by attaching
yoto an utterance, the speaker in effect indicates that his utterance serves to resolve some choice facing the hearer which the context prior to the utterance was not sufficient to resolve.
The semantics ofyopresented in this section rather directly captures the use ofyoin examples like (1). In brief, (1) presents a background context in which A is trying to decide whether to eat before going to the movies. The assertion by B places a new fact into the common ground.
2We can further articulate the context by having a potentially different set of actions for each discourse participantx, Ax(c).
3Note that the set of actions is defined relative to theoutputcontext. In example (1), where the alternative actions have already been made salient in the prior context, this set will be the same in both the input and output contexts. But examples like the following (due to an anonymous reviewer) show thatyocan be used to indicate actions that were not salient in the prior discourse context:
Context: The speaker notices that the hearer has dropped his ticket, a fact that the hearer does not realize.
a, ah,
kippu-o ticket-acc
otoshimashita dropped
yo yo
“Oh, you dropped your ticketyo.”
Withoutyo, the sentence is infelicitous; native speakers report that it sounds as if B is “just stating facts”, without expressing any connection between what he is saying and the problem faced by A. By usingyo, B indicates that the post-update context in which his assertion has been integrated is sufficient to resolve A’s problem. This invites an inference as to why the post-update context resolves A’s decision problem, and in what direction.
In the next section, I adduce quantitative support for the analysis presented in this section, relying on data from a large Japanese sentiment corpus. As will be seen, the decision-theoretic analysis developed on the basis of hand-crafted examples like (1) receives further support from the distribution ofyoin the corpus.
3 yoand Speaker Sentiment: Evidence from Sentiment Corpora
The data in this section come from a recently expanded version of the publicly available UMass Amherst Sentiment Corpora (Constantet al., 2009). The Japanese portion of this corpus contains approximately 33 million words of review text culled from reviews of various products (books, dvds, electronics, and games) appearing on the Japanese Amazon website, Amazon.co.jp. All reviews on the site are associated with a product rating given by the reviewer, ranging from 1 to 5 stars. The ratings data provide an objective scale along which the author’ssentimentorevaluation of the target product can be estimated. 1 and 5 star reviews are extremely negative and positive, respectively, while 2 and 4 star reviews are associated with more moderate negative and positive evaluations. 3 star reviews are associated with a high degree of ambivalence or lack of a strong evaluative stance with respect to the target product.
To analyze the association between specific lexical items and associated rating scores, the rel-ative frequencyof an item across the five rating categories is calculated.4The rating categories are transformed to asentiment indexsuch that sentiment index = star rating−3, so that a star rating of 3 maps to a centered sentiment index of 0 on thex-axis of the graphs to be presented. In this way, negative numbers reflect negative evaluations (1 and 2 star reviews correspond to sentiment indices of −2 and−1), and positive numbers reflect positive evaluations (4 and 5 star reviews correspond to sentiment indices of1and2).
Figure 1 shows the distribution of the English expressiveswow and damn along with yo in the review texts of the English and Japanese Amazon corpora across the five centered ratings categories. The y-axis plots the log oddsof the item in the corpora. The use of log odds allows us to fit logistic regression models to the data, in order to test for the statistical significance of certain trends in the distribution of an item across rating categories. All three items have a clear U-shaped distribution across the rating categories, an impression that is confirmed by the significance of the quadratic terms in the associated quadratic logistic regression. The U-shaped distribution indicates a tendency for these items to be used in reviews whose author has a more extreme opinion toward the item being reviewed, with a correspondingly strong recommendation, whether positive or negative.
It is conceivable that expressives like wow or damn directly index speaker emotionality, in which case its distribution in the corpus is a direct reflection of its meaning, insofar as review cat-egory serves as a proxy for emotional state. This use of the sentiment data relies on a (potentially indirect and fuzzy) mapping from emotional state to sentiment index, and vice-versa. Their distri-bution across sentiment indices thus supports the analysis of these item as expressing heightened speaker emotionality, and at the same time provides a means for empirically estimating the degree of heightened emotion expressed by this item by comparison with other expressive items. This per-spective has been adopted for the analysis of expressive items exhibiting a U-shaped distribution in these corpora (Potts and Schwarz, 2008; Constantet al., 2008).
4
context. The salient action is naturally interpreted as ordering sea urchin. The second utterance is made in the context generated, in part, by the first utterance. Usingyoin the second utterance commits the speaker to the existence of some action that is salient and optimal in the new post-update context, but not optimal in the input context. But since the second utterance is about sea urchin, the most natural interpretation is that the utterance is still suggesting that the hearer buy sea urchin. But this was already an optimal action in the input context, due to the use ofyoin the prior utterance. So the presupposition ofyois not satisfied, and the second utterance is infelicitous withyo.
These facts support the decision-theoretic analysis account ofyo’s corpus distribution. While the usage profile of yomatches a canonical expressive likewowordamn, non-repeatability pro-vides some reason to think that this profile is generated in a distinct way. For both types of items, we see a systematic distributional effect in our corpus, but the explanation for that effect is differ-ent. Expressive likedamnindex speaker emotionality directly, and rating category correlates (by hypothesis) with this emotional index. The particleyo, by contrast, serves as a guide to optimal action, and this is also reflected systematically in the rating category of the review.
In the next section of the paper, I explore one further aspect of the decision-theoretic account ofyo: it’s context-dependency. I show thatyo tends to be used later in the text of a review, and suggest how this fact might arise from the semantics given to the particle in this paper.
5 Sentence Final, Discourse Final
The particleyois syntactically restricted to matrix clause-final position. Examination of the corpus data shows a tendency foryoto appear text-finally as well. In this subsection, I present statistical evidence from the sentiment corpus supporting this generalization. I then discuss the way in which this empirical generalization fits within the theory ofyooutlined above.
To explore the textual position ofyo, I extracted from the Japanese Amazon corpus every review containing one or more instances of a matrix, sentence-final use ofyo. This excludes uses ofyoin quotative contexts, as well as cases whereyois followed by another particle; such cases do not fall within the analysis presented in this paper.6 A total of 4,486 reviews were found containing such tokens ofyo, containing a total of 5,283 tokens. The textual position of each token ofyowas then calculated by counting the number of characters that precededyoin the text. For a given review text, we can then get the textual position ofyoby dividing the textual position ofyoby the total number of characters in the text, to get a value between 0 and 1.7
The sentence-finality of yo introduces a confound in the calculation of textual position de-scribed above. To illustrate, consider a subset of reviews consisting of just two sentences of roughly equal length. Syntactically, yo can only occur at the end of the first sentence, or at the end of the second sentence. If it occurs at the end of the first sentence, its textual position will be approximately 0.5, or halfway through the text. If it occurs after the second sentence, its textual position will be 1. Ifyooccurs equally often on the first or second sentence in such reviews, then the average textual position will come out to 0.75. The sentence-finality ofyohas introduced a bias towards occurring later in the text, which has nothing to do with discourse or text-level constraints on the use ofyo.
To eliminate this confound, I calculated a corrected textual position for each occurrence ofyo
using the following procedure: I calculated the average sentence length in a review, then subtracted half of the average sentence length from the character position of each occurrence of yo in that review. In the example outlined above, this would give corrected textual positions of 0.25 for a
6In particular, the particle sequenceyo neis excluded from consideration. 7
token ofyo occurring after the first sentence, and a value of 0.75 for a token occurring after the final sentence. The corrected average textual position for a set of two-sentence reviews with an equal likelihood ofyoafter either sentence would tend toward a mean corrected position value of 0.5.
The graph in Figure 2 shows a histogram and estimated density plot of the corrected textual position ofyoin the corpus. The mean value of the corrected textual position is 0.6, with a median of 0.67. Even with the corrected positional values, it is clear that there is bias toward later positions in the text, with a highly skewed distribution of values. This distribution can be compared with that of the question particlekaand the discourse particlene, both of whose syntactic distribution is similar to that ofyo, in that they must appear sentence-finally.8 The estimated densities for these particles across textual positions were calculated using the same procedure as described for yo. The mean corrected textual position of kais 0.49, with a median value of 0.51. The mean value for neis 0.52, with a median value of 0.55. As can be seen from the graph in Figure 2, neither particle is as biased toward text-finality as yo, although ne seems to exhibit a slight bias in the same direction, for reasons I do not know.
−0.2 0.0 0.2 0.4 0.6 0.8 1.0
0.0
0.5
1.0
1.5
2.0
2.5
Corrected Textual Position of yo, Compared to ne and ka
( Position in Text − (Avg Sentence Length / 2 )) / Length of Text
Density
yo ne ka
Figure 2:Histogram and density plot showing the density of the corrected position ofyoat different points in the review text. Density estimates for two other sentence final particles are provided for comparison.
The empirical tendency for text-finality ofyofollows from its semantics when we make a few idealizations about the structure of the review texts and the rhetorical strategies adopted by authors. In the case of extremely favorable or extremely negative reviews, we can assume that most or all of the sentences in the review will be positive or negative, respectively. In the case of a 5-star review, for example, we expect a text whose sentences are uniformly positive with respect to the product. Each sentence provides a fact or sentiment that supports the conclusion that one should buy the product. The first sentence in the review is made in a null context, and adds a single fact or sentiment relevant to the question of whether to buy the product. This adds a piece of information relevant to this decision, intended to sway the reader toward buying the product. The next sentence is made in the (positive) context created by the previous sentence. If this sentence is also positive, we now have a context withtwopieces of information supporting the conclusion advocated by the
8
author. And so on. When the author uses yo, they indicate that the issue has now been settled. Rhetorically, it makes sense to save this sort of move for last.
An example will help to illustrate the principle. The text in (9) is an English translation of the review text from a 5-star review of the children’s bookHyakkai-date no ie“The One-Hundred Story House”. I have numbered the sentences to aid the discussion. The review consists of three sentences. The first sentence is neutral, describing the situation that led her to buy the book. The second sentence provides a positive sentiment. This is followed up by another sentence expressing an additional fact that supports the positive sentiment expressed by the review. This final sentence is the one that is marked byyo.
(9) [1] When I asked my 1st grade child what book he would like for reading over the summer, he answered “The One-Hundred Story House”, so I promptly searched for and bought it. [2] Watching my son reading it over and over, I felt glad for buying it. [3] My four year old daughter also listens to the story enthusiasticallyyo.
Why should yo be used to mark the last sentence in this review? I propose that the reason relies on the cumulativity of contextual update. In the context of sentence [2], we have no other positive pieces of information, while sentence [3] contains as background information the positive sentiment expressed by sentence [2]. The result is that the post-update context after sentence [3] supports the author’s conclusion to a greater degree than the post-update context after sentence [2]. The presupposition ofyorefers to the degree to which the entire post-update common ground supports a particular action, rather than the degree to which theyo-marked sentence itself supports the conclusion. The more positive sentences that have been asserted, the greater the degree to which the common ground supports the positive conclusion “buy the product”, and thus the greater the degree to which it supports the felicitous use ofyo.
The same holds in the case of negative reviews, where the more negative sentiments that have been expressed, the greater the degree to which the common ground supports the negative conclu-sion “do not buy the product”. This illustrated by the 1-star review of a video game strategy guide in (10).
(10) [1] It’s just a “dictionary” in which the data from the software has been put on paper. [2] No art, no effort; I had been looking forward to it, but was really disappointed. [3] With simple data like this, it’s easier just to check a wiki or something. [4] It would be nice if there were advice about weapons and armor, different ways to play, or strategies for difficult quests, but when I read this there was nothing interesting. [5] For people with an internet connection, there is absolutely no need for this book. [6] A book made like this is behind the timesyo.
The review consists of six sentences, all of which are highly negative. The entire review builds a strong case for not buying the book, which is emphasized by the final use of yo. Notice that the information provided by sentence [6], on which yo occurs, does not seem to be a stronger strike against the book than any of the other negative sentences [1-5]. The fact thatyooccurs with this sentence is not because it expresses a more negative or powerful argument against purchasing the book than the other sentences. Instead, its text final position follows from the fact that it is here that the argument against buying the product is strongest, since it contains all of the negative information in sentences [1-6].
information to favor one action over another. Looking at things from the other direction, once an author has usedyo, he has rhetorically indicated that he takes the issue to be settled. Such an issue-settling move, I suggest, tends to be made text-finally.
6 Conclusion
In this paper, I used data from a large sentiment corpus to explore a decision-theoretic account of the Japanese discourse particleyo. I showed how the structure of sentiment corpora can be mapped onto a decision-theoretic model of discourse contexts, and argued that this structure is consistent with a decision-theoretic account of the particle. The distribution is, however, consistent with an expressive analysis as well, so that multiple lines of evidence are needed in order to get at the right account. Two additional pieces of evidence were adduced to this end. The non-repeatability ofyowas explained in terms of the decision-theoretic account, as was the tendency toward text-finality in the sentiment corpus. I hope to have shown how data from sentiment corpora can be combined with other data in developing decision-theoretic models of meaning on a firm empirical foundation.
References
Benz, A., G. J¨ager, and R. van Rooy. 2005. Game Theory and Pragmatics. Palgrave Macmillan.
Benz, A., G. J¨ager, and R. van Rooy. 2005 An Introduction to Game Theory for Linguists, chap-ter 1. Game Theory and Pragmatics. Palgrave MacMillan, Houndsmills, Basingtoke, Hamp-shire.
Constant, N., C. Davis, C. Potts, and F. Schwarz. 2009. The pragmatics of expressive content: Evidence from large corpora. Sprache und Datenverarbeitung33(1-2).
Constant, N., C. Davis, C. Potts, and F. Schwarz. 2009. UMass Amherst sentiment corpora.
Davis, C. 2009. Decisions, dynamics, and the Japanese particle yo. Journal of Semantics 26(4):329-366.
Davis, C. and C. Potts. To appear. Affective demonstratives and the division of pragmatic labor. InProceedings of the 17th Amsterdam Colloquium.
Kratzer, A. 1981. The notional category of modality. In Words, Worlds, and Contexts. New Approaches in Word Semantics, pages 38–74. de Gruyter, Berlin.
Liu, B. 2010. Sentiment Analysis and Subjectivity. Handbook of Natural Language Processing, Second Edition.
Parikh, P. 2001. The Use of Language. CSLI, Stanford.
Portner, P. 2007. Imperatives and modals. Natural Language Semantics, 15(4):351–383.
Potts, C. 2007. The expressive dimension. Theoretical Linguistics, 33(2):165–197.
Potts, C. and F Schwarz. 2008. Exclamatives and heightened emotion: Extracting pragmatic generalizations from large corpora. Ms., UMass Amherst.
Stalnaker, R. 1978. Assertion. Syntax and Semantics, 9:315–332.
van Rooy, R. 2003. Questioning to resolve decision problems. Linguistics and Philosophy, 26:727–763.
